*
Brij Masand, HPCwire: Robert Grossman discusses managing, mining
large data sets Publications: *
GPS, First Issue of DMKD journal is available on-line in PDF
format,
--
Data Mining and Knowledge Discovery community, focusing on the
latest research and applications.
Submissions are most welcome and should be emailed, with a
DESCRIPTIVE subject line (and a URL) to gps.
Submissions may be edited for length.
Please keep CFP and meetings announcements short and provide
a URL for details.
KD Nuggets frequency is 3-4 times a month.
Back issues of KD Nuggets, a catalog of data mining tools
('Siftware'), pointers to Data Mining Companies, Relevant Websites,
Meetings, and more is available at Knowledge Discovery Mine site
at
********************* Official disclaimer ***************************
All opinions expressed herein are those of the contributors and not
necessarily of their respective employers (or of KD Nuggets)
*********************************************************************
~~~~~~~~~~~~ Quotable Quote ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
'When you come to a fork in the road, take it.'
- Yogi Berra - Previous1NextTop
Date: Thu, 15 May 1997 22:22:53 -0700
From: Ronny Kohavi (ronnyk@starry.engr.sgi.com)
Subject: Silicon Graphics' MineSet used in Incyte's LifeTools 3D
A recent press release by Incyte Pharmaceuticals Inc. announces
LifeTools 3D, a powerful data mining and visualization software based
on Silicon Graphics' MineSet(tm) software suite of data analysis and
visualization tools. In collaboration with Silicon Graphics, Incyte
created customized functions that are specifically designed to help
researchers view, explore, and identify novel genes within LifeSeq.
Engineering Manager, Analytical Data Mining. Previous2NextTop
From: 'Prof. Zicari' (zicari@informatik.uni-frankfurt.de)
Date: Fri, 9 May 1997 23:39:14 +0200 (METDST)
Subject: COMDEX Internet Application Awards.
News Release
First COMDEX Internet Application Awards
IBM, Microsoft and SUN to sponsor Awards Program for the new generation of
Internet applications
Frankfurt -- April 1997. The three leading IT companies IBM, MICROSOFT and
SUN Microsystems will jointly support an international Awards Program
designed for the new generation of Internet-based applications for
business.
The first COMDEX Internet Application Awards will be given out in the
following three categories:
Best Intranet-based application for enterprise usage
Focus: Use of an Intranet for Institutional/Corporate knowledge
for competitive advantage.
Most Innovative Web Site
Focus: Best or most innovative Web Site with respect to user
interface, easy to use, innovative content.
Best Transactional Internet Application
Focus: Database, interactive applications.
The Award winners will be selected among the submittals by a jury of
international experts. The Awards ceremony will take place on October 8,
1997 at the trade show COMDEX Internet & Object World Frankfurt'97 (October
7-10,1997, Sheraton Conference Center, Frankfurt/Main Airport).
'Successful Internet technologies like Java confirm us in considering the
Internet as the future base for enterprise computing. The COMDEX Internet
Application Awards program provides an excellent forum for honoring and
supporting outstanding Internet applications. We are looking forward to an
exciting contest', says Gert Haas, Marketing Director, SUN Microsystems,
Germany.
Microsoft's commitment to the Awards Program is explained by Karl-Heinz
Breitenbach, Customer Unit Manager Internet & Developer Customer Unit,
Microsoft Germany: 'The availability of all relevant information at work is
the base for a fast and successful decision in a company. We therefore have
taken the challenge of providing 'information at your fingertips' very
early and this is reflected by our current product line. Internet
technology today allows to rapidly and reliably represent information
distributed in all branches of the company via a so called Intranet
solution. With the sponsorship of the COMDEX Internet Application Awards,
Microsoft confirms its commitment to innovative Internet technologies which
perfectly match our company goals.'
Sanyaya Addanki, General Manager of Network Computing Solutions, IBM EMEA,
explains IBM's motivation for a sponsorship: 'IBM is committed to providing
companies with solutions that link business critical applications and data
with the global reach and easy access of the web. We are proud to sponsor
the COMDEX Internet Application Awards Program, which fosters the
development of electronic business applications. Electronic business is the
cornerstone of IBM's network computing vision.'
send an e-mail to: LogOn@omg.org
call LogOn at: +49-6173-9558-51
COMDEX Internet and Object World Frankfurt '97 are produced by SOFTBANK
COMDEX Inc. and LogOn Technology Transfer GmbH.
The show is sponsored by: Object Management Group (OMG), A1-Solutions,
Business Online, Computer Associates, Computer Zeitung, MID and redmond's.
Internet and Wireless are sponsored by Omnilink Internet Service Center and
ARtem.
Information on Conferences and Exhibition:
Christiane Sattler
LogOn Technology Transfer GmbH
Burgweg 14, D-61476 Kronberg/Ts., Germany
phone: +49-6173-9558-53
fax: +49-6173-9404-20
e-mail: LogOn@omg.org
Web:
Previous3NextTop
[the following article is included with the permission of HPCwire. GPS]
Date: Fri, 23 May 1997 14:00:49 -0400
From: Brij Masand (brij@gte.com)
Subject: ROBERT GROSSMAN DISCUSSES MANAGING, MINING LARGE DATA SETS
[From H P C w i r e *** May 23, 1997: Vol. 6, No. 20 ***]
ROBERT GROSSMAN DISCUSSES MANAGING, MINING LARGE DATA SETS
by Alan Beck, editor in chief, HPCwire 05.23.97
=============================================================================
Chicago, Ill. -- Issues raised in the effective archiving, managing and
mining of very large data sets have significant pragmatic repercussions
throughout both commercial and scientific computing. To learn more about the
state of the art in this area, HPCwire interviewed Robert Grossman, professor
of mathematics, statistics and computer science at the University of Illinois
at Chicago, president of Magnify, and principal researcher in the Terabyte
Challenge.
-------------------
HPCwire: Please give an overview of the current status of the Terabyte
Challenge, including funding sources and participants.
GROSSMAN: 'The Terabyte Challenge is open, distributed test bed for
managing and mining massive data sets. The infrastructure for the Terabyte
Challenge is provided by the NSF sponsored National Scalable Cluster Project
(NSCP) and its industrial partners. The NSCP philosophy is to use commodity
components with high performance networking to build virtural platforms with
supercomputing power. The software tools developed for the Terabyte Challenge
seek to balance high performance computing with the high performance
input/output required by data intensive and data mining applications.
'Currently, the NSCP consists of approximately 25 nodes and 500 Gigabytes
of disk at both UIC and UPenn, together with smaller clusters at the
participating partners. The infrastructure will be more than doubling over
the next few months to over 100 nodes and 2 Terabytes of disk. Unlike
other centers, the NSCP is configured for managing and mining large data
sets, ranging in size from 100 to 500 Gigabytes.
'We are currently planning the third Annual Terabyte Challenge, which
will take place at SC 97. The first two took place at Supercomputing 95 and
96 (both won High Performance Computing Challenge Awards).
'Currently, the University of Illinois at Chicago, the University of
Pennsylvania, and the University of Maryland form the core academic team. Two
industrial partners-HUBS (Philadelphia) and Magnify, Inc. (Chicago) will also
be working closely on this year's Terabyte Challenge. Funding is provided by
NSF to the NSCP Consortium, by DOE to UIC and UPenn, and by DOD to Magnify.
We expect additional partners to join us. If interested, please contact RLG.
'Current applications include mining scientific data (UIC and UPenn),
mining medical data (UIC and UPenn), detecting network intrusions with data
mining (Magnify, Inc), and data intensive computing in support of virtual
reality (HUBS).
HPCwire: What progress has been made in scaling algorithms for very large
data sets?
GROSSMAN: 'I use the 10x rule: one can expect to archive 10-100x more data
than one can manage, and manage 10-100x more data than one can mine. This
makes sense since archiving requires a simple retrieval of files or objects,
managing requires the ability to perform simple queries, and mining requires
statistically and numerically intensive queries. At SC 96, we mined data sets
that were roughly 100-250 Gigabytes in size using 10-25 nodes. At SC 97, we
hope to mine 500-1000 Gigabytes of data on 50-100 nodes. I want to emphasize
that one can manage and perform simple queries of much larger data sets (up
to tens of Terabytes), but the detailed data mining of even a few hundred
gigabytes of data is a challenge today.'
'Parallelizing data mining algorithms can be done in several ways. Most
data mining algorithms are sufficiently compute-intensive that they work best
when the data and the working space required for the algorithm fit into
memory. For large data sets this is not clearly not possible and the
challenge is to balance the i/o requirements of the algorithm with the cpu
requirements. Several approaches are possible:
'For the purposes here, we assume that the data mining process consists of
several steps, including 1) extracting patterns, 2) using these patterns
automatically to build predictive models, and 3) selecting or combining
multiple predictive models to produce a single decision. In each of the four
methods described next, one or more subsets of the data are chosen and mined.
The methods differ in how the subsets are chosen: the subsets may be created
by random draws, by a partition of the data, by a cover of the data, or by a
range based query of the data.
'In sample based data mining, one samples a large data set and then
extracts a patterns or builds a model. This is the most common approach. It
works well for patterns that are still easily found after down sampling. It
has the advantage that the compute time is vastly reduced (since the data to
be mined is vastly smaller) and the disadvantage that the patterns obtained
are often not indicative of the whole data set -- this is closely related to
the problem of over-fitting. This approach is most often not parallelized,
although sometimes sampling can be done in parallel and the results combined
into one model using model averaging techniques.
'In partitioned based data mining, the data set is partitioned into distinct
subsets which fit into memory, each partition is separately mined to produce
a collection of predictive models, and then the predictive models are
combined using model selection and model averaging techniques. This type of
data mining is easily parallelized, since one (or more) processors can be
assigned to each partition.
'Cover-based data mining is similar to partitioned based data mining, but
the different subsets to be mined can be overlapping. This is closely
related to what is called local mining, in which the patterns extracted use
data which is localized in some fashion, say based on the N closest data
points to a fixed reference point.
'Attribute-based data mining creates different subsets to be mined by using
an attribute based query of the underlying data set. For example, all objects
whose first attribute is less than 1.1 and whose second attribute is equal to
'A', etc. are selected and then mined.
'For more information, see R. L. Grossman, Scaling Data Mining Algorithms
Using Cover-based Learning with Model Selection and Model Averaging,
HPCwire: How is the TC approaching the mining of highly distributed data?
GROSSMAN: 'On the systems side, we have made good progress in this area.
The NSCP clusters at UIC and UPenn have been connected for several weeks now
by the vBNS at OC-3 (155 Mbps) speeds. Using this infrastructure we have
experimented with wide area data mining of scientific and medical data. We
are currently using this experience to develop new algorithms for wide area
data mining and to develop new generations of our data management and data
mining tools. The challenge is to develop a new class of algorithms for
extracting patterns from widely distributed data without the necessity of
first warehousing the data.'
HPCwire: What progress has been made in better understanding dynamical
systems via data mining?
GROSSMAN: 'Not as much as we would have liked. Data mining algorithms
today, by and large, work with data which is flat and static. The core
dynamical system concepts of a state vector and its evolution in time are
missing in most data mining algorithms. Hybrid systems is an emerging field
which combines dynamical systems with discrete structures such as rule
systems and automata. The latter can express the patterns discovered in data
mining. Researchers working in the NSCP are actively investigating exploiting
hybrid systems and related techniques to develop next generation data mining
algorithms which can utilize state information and work with time varying
data.'
HPCwire: How is TC research being made available to the commercial sector?
Have any new products or partnerships resulted from TC-generated technology?
GROSSMAN: 'The NSCP and the Terabyte Challenge have 1) published the core
ideas they have developed for data mining and data intensive computing, 2)
developed reference architectures and implementations for software tools to
support data mining (the UIC software tools PTool, JTool, and DMTool), and 3)
encouraged companies to exploit this technology for data intensive computing
and data mining.
'To date, HUBS in Philadelphia and Magnify, Inc. in Chicago have begun to
employ some of these ideas in the products and services they offer.
Currently, regional data minings centers are in the planning process in both
Chicago and Philadelphia.'
HPCwire: How do you see the TC evolving over the next five years?
GROSSMAN: 'The most exciting development is the expected transformation of
the NSCP into two regional data mining centers with very strong industrial
ties: one in Chicago and one in Philadelphia. This has three important
consequences: 1) First the compute, i/o, and networking infrastructure which
we can dedicate to data mining projects is expected to double this year and
hopefully to double again in about two years. 2) With our industrial
partners, we are actively working to demonstrate the practical feasibility of
mining massive data sets and to establish open standards for managing,
mining, and modeling massive data sets. 3) Using the vBNS network connecting
the centers in Chicago and Philadelphia, we are finding it easy to experiment
with the type of wide area data mining issues which we expect to take on an
increasing important role for scientific, engineering, medical, and business
data mining applications.
'To summarize, during the next five years, we expect the Terabyte Challenge
not only to continue to push the boundaries of massive data mining through an
annual competition, but also, together with its industrial partners, to be
actively involved with establishing data mining standards and reference
implementations of software tools for managing, mining, and modeling massive
data sets.
'Additional participants for 1997 competition are welcome. Please contact
one of the organizers if interested. Additional information
can be found at
--------------------
Alan Beck is editor in chief of HPCwire. Comments are always welcome and
should be directed to editor@hpcwire.tgc.com
Copyright 1997 HPCwire. Redistribution of this article is forbidden by law
without the expressed written consent of the publisher. For a free trial
subscription to HPCwire, send e-mail to trial@hpcwire.tgc.com.
H P C w i r e
The Text-on-Demand E-zine for High Performance Computing
***************************************************************************
Previous4NextTop
Date: Thu, 22 May 1997 15:05:54 -0400
From: Gregory Piatetsky-Shapiro (gps)
Subject: First Issue of DMKD journal is available on-line in PDF format
The premiere issue of Data Mining and Knowledge Discovery journal
is available on-line, in PDF format, at
Only the first issue will be freely available on-line,
but you can subscribe to the journal for $50 individual rate, more
for institutional rate
-- see
for subscription information. Please support this journal !
Previous5NextTop
From: A.N.Pryke@cs.bham.ac.uk
Date: Fri, 23 May 97 22:12:09 BST
Subject: Nuggets: Bibliography of KDD and Data Mining Papers
The Master Bibliography of KDD and Data Mining Papers is a
bibliography of over 400 papers on the topics of Data Mining and
Knowledge Discovery in Databases (this includes closely related papers
on visualisation and machine learning). More than 70 of the papers are
online.
It is available in either bibtex, or html annotated bibtex formats
from:
Andy additional references, or corrections are gratefully
received. Please email them to me, Andy Pryke, at
A.N.Pryke@cs.bham.ac.uk
Only references in machine readable format
(e.g. refer or preferable Bibtex) can be added, due to time
constraints.
Note that all the information I have about the papers in in the
bibliography, and many (330ish) of the papers are not available
online.
Please read the _collection_ copyright statement at
Previous6NextTop
Date: Fri, 16 May 1997 19:09:05 -0500
From: dfisher@vuse.vanderbilt.edu
(Douglas H. Fisher)
Subject: COLT/ICML Early Registration
Early registration for the Tenth Annual Conference on
Computational Learning Theory (COLT-97) and/or the Fourteenth
International Conference on Machine Learning (ICML-97)
concludes June 2, 1997. Room blocks at area hotels and on campus
are also 'released' June 2 (though rooms will likely still be available
after that date). See
Previous7NextTop
Date: Fri, 16 May 1997 16:44:56 +0200 (MET DST)
From: Jan Komorowski (Jan.Komorowski@idi.ntnu.no)
Subject: PKDD'97 -- Call For Participation
1st European Symposium on Principles of
Data Mining and Knowledge Discovery in Databases
Trondheim, Norway
June 24-27, 1997
Tutorials: June 24-25
Symposium: June 26-27
This is an invitation to the 1st European Symposium on Principles of
Data Mining and Knowledge Discovery in Databases.
PKDD'97 is the first symposium in an intended series of meetings of
the data mining and knowledge discovery from databases (KDD) community
in Europe. The goal of the PKDD series is to provide a European-based
forum for interaction among all theoreticians and practitioners
interested in data mining and knowledge discovery. Fostering an
interdisciplinary collaboration is one desired outcome, but the main
long-term focus is on theoretical principles for the emerging
discipline of KDD, especially those new principles that go beyond each
of the contributing areas.
There were 50 papers submitted to PKDD'97. After the selection by the
program committee, the papers were assigned into three categories: 14
plenary papers, 13 parallel session papers and 11 poster papers that
include spot-light presentations in the plenary sessions. In
addition, four tutorials were selected: Rough Sets for Data Mining and
Knowledge Discovery, Techniques and Applications of KDD, High
Performance Data Mining, and Data Mining in the Telecommunications
Industry.
The proceedings are published by Springer Verlag.
The invited speakers include Evangelos Simoudis, USA, and Bjarne Foss,
Norway. Theey will provide their different perspectives on the field:
one is data mining for businesses and the other data mining seen from
the point of view of control theory. Panel discussions on the present
situation and the future development of the field are planned.
There will be software exhibitions of both commercial and academic
software.
for
detailed information and news about the symposium.
Previous8NextTop
From: David Heckerman (heckerma@MICROSOFT.com)
Subject: Summer School on PROBABILISTIC GRAPHICAL MODELS
Date: Fri, 16 May 1997 08:08:00 -0700
A Newton Institute EC Summer School
PROBABILISTIC GRAPHICAL MODELS
1 - 5 September 1997
Isaac Newton Institute, Cambridge, U.K.
Organisers: C M Bishop (Aston) and J Whittaker (Lancaster)
Probabilistic graphical models provide a very general framework for
representing complex probability distributions over sets of
variables. A powerful feature of the graphical model viewpoint is that
it unifies many of the common techniques used in pattern recognition
and machine learning including neural networks, latent variable
models, probabilistic expert systems, Boltzmann machines and Bayesian
belief networks. Indeed, the increasing interactions between the
neural computing and graphical modelling communities have resulted in
a number of powerful new ideas and techniques. The conference will
include several tutorial presentations on key topics as well as
advanced research talks.
Provisional themes:
Conditional independence; Bayesian belief networks; message
propagation; latent variable models; variational techniques; mean
field theory; learning and estimation; model search; EM and MCMC
algorithms; axiomatic approaches; causality; decision theory; neural
networks; information and coding theory; scientific applications and
examples.
Provisional list of speakers:
C M Bishop (Aston) D J C MacKay (Cambridge)
R Cowell (City) J Pearl (UCLA)
A P Dawid (UCL) M D Perlman (Washington)
D Geiger (Technion) M Piccioni (Aquila)
E George (Texas) R Shachter (Stanford)
W Gilks (Cambridge) J Q Smith (Warwick)
D Heckermann (Microsoft) M Studeny (Prague)
G E Hinton (Toronto) M Titterington (Glasgow)
T Jaakkola (UCSC) J Whittaker (Lancaster)
M I Jordan (MIT) S Lauritzen (Aalborg)
B Kappen (Nijmegen) D Spiegelhalter (Cambridge)
M Kearns (AT&T) S Russell (Berkeley)
This instructional conference will form a component of the Newton
Institute programme on Neural Networks and Machine Learning, organised
by C M Bishop, D Haussler, G E Hinton, M Niranjan and L G Valiant.
Further information about the programme is available via the WWW at
The conference will take place in the Isaac Newton Institute and
accommodation for participants will be provided at Wolfson Court,
adjacent to the Institute. The conference package costs 270 UK pounds
which includes accommodation from Sunday 31 October to Friday 5
September, together with breakfast, lunch during the days that the
lectures take place and evening meals.
Applications:
To participate in the conference, please complete and
return an application form and, for students and postdoctoral fellows,
arrange for a letter of reference from a senior scientist. Limited
financial support is available for participants from appropriate
countries.
Application forms are available from the conference Web Page at
Completed forms and letters of recommendation should be sent to Heather
Dawson at the Newton Institute, or by e-mail to
h.dawson@newton.cam.ac.uk
*Closing Date for the receipt of applications and
letters of recommendation is 16 June 1997*
Previous9NextTop
From: Vasant Honavar (honavar@cs.iastate.edu)
Subject: Call for Participation: Workshop on Automata Induction,
Grammatical Inference, and Language Acquisition
Date: Thu, 8 May 1997 10:53:48 -0500 (CDT)
Workshop on
Automata Induction, Grammatical Inference, and Language Acquisition
The Fourteenth International Conference on Machine Learning (ICML-97)
July 12, 1997, Nashville, Tennessee
The Automata Induction, Grammatical Inference, and Language Acquisition
Workshop will be held on Saturday, July 12, 1997 during the Fourteenth
International Conference on Machine Learning (ICML-97) which will be
co-located with the Tenth Annual Conference on Computational Learning Theory
(COLT-97) at Nashville, Tennessee from July 8 through July 12, 1997.
Additional information on ICML-97 and COLT-97 can be found at
Previous10NextTop
Date: Wed, 21 May 1997 12:23:13 +1000
From: Honghua Dai (dai@cs.monash.edu.au)
Subject: KDEX-97 Final Call for Papers
1997 IEEE Knowledge and Data Engineering Exchange Workshop (KDEX-97)
--------------------------------------------------------------------
Sponsored by the IEEE Computer Society and Co-located with
the 9th IEEE Tools with Artificial Intelligence Conference
November 4, 1997, Newport Beach, California, U.S.A.
===================================================
Call for Papers
The 1997 IEEE Knowledge and Data Engineering Exchange Workshop
(KDEX-97) will provide an international forum for researchers,
educators and practitioners to exchange and evaluate information and
experiences related to state-of-the-art issues and trends in the areas
of artificial intelligence and databases. The goal of this workshop
is to expedite technology transfer from researchers to practitioners,
to assess the impact of emerging technologies on current research
directions, and to identify emerging research opportunities.
Educators will present material and techniques for effectively
transferring state-of-the-art knowledge and data engineering
technologies to students and professionals. The workshop is currently
scheduled for an one-day duration, but depending on the final program
it might be extended to a second day.
Submissions can be in the form of survey papers, experience reports,
and educational material to facilitate technology transfer. Accepted
papers will be published in the workshop proceedings by the IEEE
Computer Society. A selected number of the accepted papers will
possibly be expanded and revised for publication in the IEEE
Transactions on Knowledge and Data Engineering (IEEE-TKDE) and the
International Journal of Artificial Intelligence Tools. Educational
material related to papers published in the IEEE-TKDE will be posted
on the IEEE-TKDE home page.
The theme of the workshop is 'AI MEETS DATABASES'. Topics of interest
include, but are not limited to:
- Computer supported cooperative processing and interoperable
systems
- Data sharing, data warehousing and meta-data management
- Distributed intelligent mediators and agents
- Distributed object management
- Dynamic knowledge
- Evaluation and measurement of knowledge and database systems
- High-performance issues (including architectures, knowledge
representation techniques, inference mechanisms, algorithms and
integration methods)
- Information structures and interaction
- Intelligent search, data mining and content-based retrieval
- Knowledge and data engineering systems
- Quality assurance for knowledge and data engineering systems
(correctness, reliability, security, survivability and
performance)
- Software re-engineering and intelligent software information
systems
- Spatio-temporal, active, mobile and multimedia data
- Emerging applications (biomedical systems, decision support,
geographical databases, Internet technologies and applications,
digital libraries, etc.)
All submissions should be limited to a maximum of 5,000 words. Six
hardcopies should be forwarded to the following address.
Xindong Wu (KDEX-97)
Department of Software Development
Monash University
900 Dandenong Road
Caulfield East, Melbourne 3145
Australia
Please include a cover page containing the title, authors (names,
postal and email addresses, telephone and fax numbers), and an
abstract. This cover page must accompany the paper.
************ I m p o r t a n t D a t e s *****************
* 6 copies of full papers received by: June 15, 1997 *
* acceptance/rejection notices: July 31, 1997 *
* final camera-readies due by: August 31, 1997 *
* workshop: November 4, 1997 *
************************************************************
Previous11NextTop
From: gordon@AIC.NRL.Navy.Mil
Date: Tue, 20 May 97 10:30:38 EDT
Subject: CFP: ICML-97 workshop on REINFORCEMENT LEARNING: TO MODEL OR
NOT TO MODEL, THAT IS THE QUESTION
Workshop at the Fourteenth
International Conference on Machine
Learning (ICML-97)
Vanderbilt University, Nashville, TN
July 12, 1997
www.cs.cmu.edu/~ggordon/ml97ws
Recently there has been some disagreement in the reinforcement
learning community about whether finding a good control policy
is helped or hindered by learning a model of the system to be
controlled. Recent reinforcement learning successes
(Tesauro's TD-gammon, Crites' elevator control, Zhang and
Dietterich's space-shuttle scheduling) have all been in
domains where a human-specified model of the target system was
known in advance, and have all made substantial use of the
model. On the other hand, there have been real robot systems
which learned tasks either by model-free methods or via
learned models. The debate has been exacerbated by the lack
of fully-satisfactory algorithms on either side for
comparison.
Topics for discussion include (but are not limited to)
o Case studies in which a learned model either contributed to
or detracted from the solution of a control problem. In
particular, does one method have better data efficiency?
Time efficiency? Space requirements? Final control
performance? Scaling behavior?
o Computational techniques for finding a good policy, given a
model from a particular class -- that is, what are good
planning algorithms for each class of models?
o Approximation results of the form: if the real system is in
class A, and we approximate it by a model from class B, we
are guaranteed to get 'good' results as long as we have
'sufficient' data.
o Equivalences between techniques of the two sorts: for
example, if we learn a policy of type A by direct method B,
it is equivalent to learning a model of type C and computing
its optimal controller.
o How to take advantage of uncertainty estimates in a learned
model.
o Direct algorithms combine their knowledge of the dynamics and
the goals into a single object, the policy. Thus, they may
have more difficulty than indirect methods if the goals change
(the 'lifelong learning' question). Is this an essential
difficulty?
o Does the need for an online or incremental algorithm interact
with the choice of direct or indirect methods?
full information at
www.cs.cmu.edu/~ggordon/ml97ws
Contact: Geoff Gordon (ggordon@cs.cmu.edu)